Building a Conversational AI Medium to Enhance Psychotherapy Training with Virtual Patients


<!-- metadata commented in wiki content


<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<big>'''Building a Conversational AI Medium to Enhance Psychotherapy Training with Virtual Patients'''</big></div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<big>Michael Petrizzo</big></div>
-->

==ABSTRACT==

Psychotherapists in training lack a standardized and formalized method of patient interaction for the proper development of empathy, communication, and experience. Currently, the experience of a resident may include training with other residents where one patient acts as the patient and one the psychotherapist, or with the usage of a simulated patient, an actor, to practice with the resident. Both methods have shortcomings in availability, reliability, and the accuracy of the patient in replicating a real scenario. This project attempted to create virtual patients by utilizing online patient transcripts through the fine-tuning and transfer-learning of three modern Artificial Intelligence models, ChatGPT-4o, LlaMa-3.1v-405B, and Gemini 1.5 Pro; alongside the miniature versions of these models as applicable. This included the creation of a website interface that can interact with the created models for evaluation, while also allowing an interactive format with a simplistic design. The accuracy of the models was independently evaluated through cosine similarities between data and model outputs to find semantic relations, and varied from 92.99% to 83.40%; with ChatGPT-4o Mini and Full having the highest fine-tuned and transfer-learning accuracies respectfully. Furthermore, a customizable model allowed user input for specific descriptions and mental illnesses. This highlights the potential for a model to be successfully representative of a patient which can be utilized to train residents easier than currently available methods. The need for further evaluation and continual training are at the forefront of current limitations.

==INTRODUCTION==

Psychotherapists, the umbrella category encompassing social workers, psychologists, psychiatrists, and general therapists, are vital supportive assets for the mental well-being of many. The training of a psychotherapist is standardized under the Accreditation Council for Graduate Medical Education [1], which requires generalized competencies in a variety of skills. To become a licensed psychotherapist, the requirements vary from state to state but generally require at least 60 hours of semester education [2].

Simulated patients/standardized patients have been used to both teach and assess residents [3] , requiring an actor trained for a specific situation but being an effective and meaningful method of education [4]. Simulated patients must themselves go through training to maintain a level of quality and standard between patients [5]. Simulated patients are also expected to give a level of interpersonal connection and have memorized events/encounters.

On the other hand, virtual patients, which are computer-simulated, combine the advantage of patient usage and eliminate the need for an actor. This broad term encompasses from decision-tree-style conversations to interactive virtual avatars. Virtual patients have been utilized for general medical practices, such as performing examinations and patient monitoring [6], while their usage as a psychotherapist training tool has been severely under-explored. Currently, the creation of virtual patients for psychotherapists relies heavily on recording simulated patients and then giving the user a limited array of options [7].  This lacks interactivity typical of patient interaction by forcing the user to select an option that may not be reflective of their response. Hence, generative methods for virtual patients may be highly beneficial as a training tool.

The objective of this study is to explore the potential of machine learning models, specifically ChatGPT-4o, Llama-3.1v-405B, and Gemini 1.5 Pro, in replicating virtual patients for psychotherapist training. The study aims to evaluate the effectiveness of these models in simulating therapeutic interactions, assess their accuracy in mimicking patient behaviors, and determine their potential for enhancing psychotherapist training by offering a scalable, interactive alternative to traditional simulated patients.

Machine Learning (ML), a rapidly advancing field, is the concept of teaching machines to perform specific tasks and detect patterns [8]. ML is sub-sectioned into different processes for differing problems, such as Recurrent Neural Networks for textual processing, Convolutional Neural Networks for image processing, and Neural Networks for general pattern processing. The concept of Machine Learning was developed under Alan Turing’s 1950 advances [9], and has been ever growing since. The training cycles of a model are entitled epochs, the adjustable aspects are hyperparameters, and often follow a methodology for tuning hyperparameters through optimization with algorithms such as AdamW (Adaptive Movement Estimation with Weight Decay) [10].  Fine-tuning is the process of using a pre-trained model which was developed on a much broader and larger dataset, and then re-training the model on a more specified dataset to develop new characteristics in the model. This retains the model’s ability for large language capability while also allowing specialization, such as a medical questionnaire or for mathematical evaluation.

Natural Language Processing (NLP), is the act of preprocessing semantic and textual data into a computer-viable format [11]. This includes lemmatization, embeddings, keyword filtering, truncation, padding, and other methods of cleansing. NLP is sequential, meaning all inputs will receive the same preprocessing techniques. NLP can be effectively split into word embedding and input filtering. Words are often broken into their core components, which processes such as lemmatization and stemming do in differing ways, either by the core word or by removing prefixes, this entire process is called tokenization. Then, these core words are embedded, which is the process of converting characters to integer lists of large dimensionality. The dimensionality maps words in space and ideally can show similarities in meaning by distance between points (Figure 1). The process of embedding is often unique to each model for various purposes, costs, and training data

[[File:Draft_Petrizzo_769196690_2990_Embedding_Relationship.png|centre|thumb|'''Figure 1''': Word Embedding Visualization (developers.google.com)]]
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;"></div>

Cosine Similarity measures the distance between dimensional vectors to generate a broad accuracy (Figure 2). Depending on the embeddings utilized, this distance measures the semantic similarity between two points, vital for measuring the similarity between two texts which are not written identically. While not providing absolute accuracies which are reliable to a complete model’s behavior, it provides an additional metric which can compare relative accuracies.

[[File:Draft_Petrizzo_769196690-image6-c.png|centre|thumb|444x444px|'''Figure 2'''<nowiki>: Cosine Similarity ( https://www.index.dev/blog/best-nlp-algorithms-to-get-document-similarity)</nowiki>]]
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<br /> Neural Networks, labeled as the core of modern ML, are a series of nodes connected via inputs and outputs in a web-like structure (Figure 3). These nodes contain activation functions that perform a rounding formula that will either activate or deactivate the node for usage [12]. The most common activation functions include ReLU, Sigmoid, Tanh, and Leaky ReLU which differ depending on the task of the model [13]. The model learns by passing training data through the neural network, and then alters the internal parameters of the node, including the activation function, to create a correct output [14]. Over the course of multiple diverse inputs and accurate outputs, the model can accurately predict the output of a new input. </div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image2-c.png|centre|thumb|348x348px|'''Figure 3'''<nowiki>: Neural Network Structure (https://www.geeksforgeeks.org/artificial-neural-networks-and-its-applications/)</nowiki>]]</div>

Recurrent Neural Networks (RNNs) are utilized predominantly in sequential data, including textual sequences [15]. RNNs utilize the fundamental neural network aspect while exhibiting “attention” or the ability to remember how a series of inputs, rather than one input, can alter an output. This is achieved by processing inputs through the neural network while also passing the unaltered inputs back into the model (Figure 4). This creates a “series” of inputs that the model can process for specific tasks.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image12-c.png|centre|thumb|326x326px|'''Figure 4''': RNN Structure.(Generated by Student)]]</div>

Transformers (TRNNs) built off the RNN structure by introducing more sophisticated attention and processing mechanisms (Figure 5). The usage of an encoder-decoder system allows the models independent processing between outputs and inputs, while also introducing customizable tokenization processes that represent the data’s features [16]. This process of introducing information back and forth between the encoder and decoder is defined through the feed-forward layers directly into a series of multi-headed attentions. The normalization and add layer serve to prevent exploding or vanishing gradients, and often “contain” the outputs being produced [17], this normalization improves training time by utilizing numbers less than 1, which exponentially decreases rather than increases. The specific architecture varies among models, but generally includes these series of multi-head attention, feed forwards, normalizations, and lastly ends with an activation function.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image1-c.png|centre|thumb|355x355px|'''Figure 5'''<nowiki>: TRNN Structure (https://www.baeldung.com/cs/rnns-transformers-nlp)</nowiki>]]</div>

Mixture of Experts (MoE) is a sub-divisional-based architecture comprising multiple networks connected similarly to TRNN to achieve self-attention [18]. This is accomplished through specialized “experts” to perform subdivided tasks of the model. Working similarly to an office, the “leader” flows all input directions, and calls upon “workers” who are the specialized experts. This type of model has lower training times and higher accuracy convergence compared to conventional models  [19]. While not a new concept, its difficult conceptualization makes it a novel approach to Machine Learning and is currently deployed in advanced modern models, such as ChatGPT 4.0.

==SCIENTIFIC FOUNDATION==

Simulated Patients are a fundamental aspect of psychotherapy training which requires a certified actor to be available. Due to this, oftentimes another psychiatric resident loses training time to serve as a simulated patient for another resident. While virtual patients seem promising as an alternative, they lack the interactivity of a simulated patient. As such, this project aimed to develop virtual patients that have the advantage of simulated patients by leveraging AI.

The advancing AI models of ChatGPT-4o, LlaMa-3.1v-405B, and Gemini 1.5 Pro have been shown to have higher levels of problem-solving and human-like characteristics [20]. Furthermore, these models have been used to mimic human behaviors, notably for medical practice with promising potential [21]. The increase in performance can best be attributed to the large parameter counts and data sizes (Table 1). Each of these models has smaller or “mini” alternatives that hold the same architecture but smaller parameters and model sizes. Ideally, the largest model can process inputs more effectively, but due to costs or training time, smaller models were utilized both for testing and to serve as a baseline comparison among models.

'''Table 1:''' Model Specifications and Parameter Comparisons. Generated by Student, Combination of Multiple Sources [22,23,24])

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;vertical-align: top;"|Model/Parameter
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;vertical-align: top;"|
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''LlaMa-3.1v-405B''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''Gemini-1.5-Pro''
|-
|  style="border-top: 1pt solid black;vertical-align: top;"|''Architecture''
|  style="border-top: 1pt solid black;vertical-align: top;"|
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|Encoder-Decoder
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|Decoder Only
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|''MoE TRNN''
|-
|  style="vertical-align: top;"|''Model Size''
|  style="vertical-align: top;"|
|  style="text-align: center;vertical-align: top;"|~500 GB
|  style="text-align: center;vertical-align: top;"|~1 TB
|  style="text-align: center;vertical-align: top;"|N/A*
|-
|  style="vertical-align: top;"|''Parameters''
|  style="vertical-align: top;"|
|  style="text-align: center;vertical-align: top;"|~1.8 Trillion
|  style="text-align: center;vertical-align: top;"|~405 Billion
|  style="text-align: center;vertical-align: top;"|N/A*
|-
|  style="vertical-align: top;"|''Speed (Tokens/second)''
|  style="vertical-align: top;"|
|  style="text-align: center;vertical-align: top;"|63.3
|  style="text-align: center;vertical-align: top;"|27.8
|  style="text-align: center;vertical-align: top;"|59.8
|-
|  rowspan='1' style="border-bottom: 1pt solid black;vertical-align: top;"|''Cost (Per Million Tokens)''


|  style="border-bottom: 1pt solid black;text-align: right;vertical-align: top;"|''Input''
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|$5.00
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Free for Research
|$1.25
|-
| style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: right;vertical-align: top;" |
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''Output''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|$15.00
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Free for Research
|$2.50
|}


* - Not disclosed publicly, but can be estimated to be within the ranges of 140B Parameters and ~250 GB.

These models were chosen for their high accuracy in token processing, problem-solving, replication, and high token density. Each model has specific strengths, which is why three models were chosen rather than a singular (Table 2). Namely, Massive Multi-Language Understanding (MMLU) tests a model’s ability to multitask among a variety of subjects (Hendrycks et al., 2020); Instruction Focused Evaluation for LLMs (IFEval) evaluates the ability of a model to follow the given instructions [25]; Zero-Shot Benchmark for Long Text Understanding (ZeroSCROLLS) assesses the ability to hop between a series of paragraphs and answer questions [26]; InfiniteBench is a questionnaire requiring an average token size larger than 100k [27]; and lastly, Needle in Haystack (NIAH), requires the model to find a fact from a large article or dataset [28]. The term N-Shot references the amount of pre-training on the data prior to testing and Chain of Thought (CoT) is a prompting style to encourage problem-solving and sequential processing [29].

Benchmarks demonstrating model efficiency among high tokenization and human-like characteristics are vital for model predictions in fine-tuning and for case-specific uses. For example, Gemini 1.5 Pro’s high token context performance allows for conversations to be continuous without attention loss and is tested among InfiniteBench, ZeroSCROLLS, and NIAH. The MMLU benchmarks show a model’s ability to be humanistic, to understand and grasp complex semantic relationships in a conversation.

'''Table 2:''' Model Comparison Among Various Key Benchmarks. Generated by Student, Combination of Multiple Sources [30,31,32,33,34,35,36]

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;vertical-align: top;"|Model/Benchmark
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''LlaMa-3.1v-405B''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''Gemini 1.5 Pro''
|-
|  style="border-top: 1pt solid black;vertical-align: top;"|''MMLU (0-shot, CoT)''
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|'''88.7'''
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|88.6
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|86.0
|-
|  style="vertical-align: top;"|''MMLU Pro (5-shot, CoT)''
|  style="text-align: center;vertical-align: top;"|74.0
|  style="text-align: center;vertical-align: top;"|73.3
|  style="text-align: center;vertical-align: top;"|'''75.8'''
|-
|  style="vertical-align: top;"|''IFEval''
|  style="text-align: center;vertical-align: top;"|85.6
|  style="text-align: center;vertical-align: top;"|'''88.6'''
|  style="text-align: center;vertical-align: top;"|87.1
|-
|  style="vertical-align: top;"|''ZeroSCROLLS/8K Alternative''
|  style="text-align: center;vertical-align: top;"|90.5
|  style="text-align: center;vertical-align: top;"|95.2
|  style="text-align: center;vertical-align: top;"|'''99.7'''
|-
|  style="vertical-align: top;"|''InfiniteBench''
|  style="text-align: center;vertical-align: top;"|82.5
|  style="text-align: center;vertical-align: top;"|'''83.4'''
|  style="text-align: center;vertical-align: top;"|-
|-
|  style="border-bottom: 1pt solid black;vertical-align: top;"|''Needle In Haystack''
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|'''100.0'''
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|98.1
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|99.7
|}

Older generation models, such as GPT 3.5 have been tested in comparative studies between human-like tendencies of models, with mixed results [37]. Although AI models trained on mental health transcripts have yet to be studied extensively, the data required for training these models is publicly available. This may be utilized to create virtual patients that have high similarities to real patients without the possible detrimental effects that can be placed on a patient by a resident in training. To ensure effectiveness, the testing of potential virtual patients among current resident training is incredibly important as well, efforts to reach individuals to test these models are ongoing. For ease of use and accessibility, the creation of a Graphical User Interface (GUI) was equally vital to successful implementation and was subject to the same amount of criticism and feedback as the model. The standardization of psychotherapy training with virtual patients can ensure proper development in the resident’s ability for patient communication, empathy, and allows easier access than compared to actors. Further, virtual patients can encompass a larger variety of mental illnesses and specific situations through the use of a singular transcription. The model being adjustable also allows the same model to have unique responses to the same input, given a change in the temperature of the model. Lastly, the integration of resident transcription allows instructors faster evaluations of the competencies and abilities of a resident.

Previous research in utilizing AI to replicate patients for training residents is novel, with very limited conceptualization and application. However, recently a new trial has been published; Patient-Ψ [38] attempts the goal of training mental health professionals via ChatGPT-4.0 specialization and prompt engineering. These prompts reflect specific emotional and situational aspects of the individual to be replicated and then use a “trainer” framework to converse between the patient and a user. The results of the model output were compared with professionals against the default ChatGPT-4.0 model. Further, this paper aims specifically at CBT training and aims for generalization in future research. Vitally, the usage of supervised learning is not explored, instead opted for unsupervised learning. Further, rapid model development warrants the usage of multiple models for analysis.

==METHODOLOGY==

=== ''Data Collection''===

Transcripts were gathered through various sources, such as Kaggle, and online publicly available PDFs through direct Google searches for “Therapy Patient Transcripts”.  In total, there are 2 transcriptions in pdf format which were then converted into Comma Separated Value (CSV) format for easier usage in eventual model training and fine-tuning. The choice of transcripts relied heavily on data quantity, sufficient patient output/communication, and data availability. Each transcript is a direct therapy session between the patient and the therapist in a one-on-one style. The conversion from PDF to CSV was hand-generated, with the assistance of ChatGPT-4.0 for demographic characteristics which included a general age range. The transcripts were formatted into full segments of patient and therapist messages, and the corresponding characteristics as columns. Independent CSV files are then combined into one large file for convenience, which are separated by index (Figure 6). The addition of further transcripts is crucial, but due to data availability and required patient anonymity, is difficult to collect without the direct guidance of a professional.

[[File:Draft_Petrizzo_769196690-image5-c.png|centre|thumb|600x600px|'''Figure 6''': Generated CSV Structure (Generated by Student)]]
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;"></div>

=== ''Interface Creation'' ===

The GUI consists of an HTML, CSS, Javascript, and Python co-creation to create a website interface and GUI simultaneously. HTML and CSS are the predominant combinations for creating professional and aesthetically appealing User Interfaces. The individual pages of the website are split into separate HTML files but utilize the same CSS style to ensure consistency. Javascript was integrated into HTML structure to allow further user input, such as custom form submissions or the creation of templates; headers, and footers which can be imported into pages and reduce redundancy. As HTML and CSS are not back-end languages, Python was utilized to connect the fundamental ML models to the front-end appearance by passing through Javascript. The model was trained in Python, which has a higher availability for ML modules, with usage in Tensorflow, Google-AI-Studio, and HuggingFace. Javascript can pass parameters between these languages through form submissions, features, functions, and queries. The individual pages are also referencing each other; allowing interaction between various components of the GUI, often with unique features that pass the user’s intended model and interaction. To create an application with the website-based code, Electron was utilized to create an environment to run the HTML, CSS, Javascript, and Python code in combination. This allows zipping into an executable format for application usage and possible Mobile conversion.

The index page or “initial” state, contains a form submission and description of model choices (Figure 7). This page serves as the connection between the user and their choice of model. The page covers the basic descriptors of each model and redirects to the fundamental “chatBot” page by passing the chosen patient as a feature. For individual model testing, further choices of ML model between GPT-4o and GPT-4o-mini are available. The usage of an account is intentionally avoided to prevent security concerns and database usage but may be added as needed to view previous conversations or create a “classroom” with a supervisor being able to view the “student’s” interactions.

[[File:Draft_Petrizzo_769196690-image8-c.png|centre|thumb|600x600px|'''Figure 7''': Main Index Page. (Generated by Student)]]
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;"> </div>

The “chatBot” page takes the chosen model and creates a unique interaction between the user and the virtual patient (Figure 8). This contains a chat box with textual or voice input, passes the input into the ML model, and returns a visual and auditory response to the user. The requests between the front-end (user interface) and back-end (python model) are sent with Flask via HTTPS Ports for local development and server-sided ports during complete implementation.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image10-c.png|centre|thumb|492x492px|'''Figure 8''': Example Conversation with Transfer-Learning GPT-4o-Mini. (Generated by Student)]]</div>

External pages may be utilized to contain contact, legal, or extraneous information.

=== ''Model Creation'' ===

Each model experienced differing training according to API availability, UI aspects, and other various preprocessing required. Llama-3.1-405B was downloaded directly onto the system, while GPT and Gemini required API usage and continual costs per input. The models utilized customized accuracy metrics of cosine similarity comparing the direct semantic relation between the model’s outputs and the patient’s characteristic behaviors, to determine the most successful model per transcription. Llama-3.1-405B and ChatGPT-4o used smaller versions for testing, specifically, Llama-3.2-1B and ChatGPT-4o mini, and were ultimately incorporated in accuracy metrics as training loss indicated high accuracy. An additional series of models were instead trained through transfer-learning, where transcriptions were instead passed through context rather than explicitly fine-tuned. For continual progression, communication with a professional is an ongoing operation ingrained into the model testing (Figure 9).

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image3-c.png|centre|thumb|492x492px|'''Figure 9''': Methodology. (Generated by Student.)]]</div>

ChatGPT-4o; This model was fine-tuned with the OpenAI API by structuring the dataset into a series of messages between the user and client. The assistant text would follow the patient portion, and the user would follow the therapist. The messages were formatted into a JSON format as {messages : [{role: system/client/user, content: text…}]} with the first dataset being BB3-Session-2-Annotated and containing 230 JSON lines. This file was uploaded into OpenAI and fine-tuned with GPT-4o-mini-2024-07-18 prior to full model use. The model was trained with 3 epochs, contained ~23558 tokens per epoch, and cost roughly ~0.01$; which was calculated through the OpenAI recommended algorithm [39]. In total, the model was fine-tuned through 690 steps, and had a final training loss of 0.1172 with a 1.00 training accuracy. The gpt-4o-full model was trained similarly, with slightly differing parameter; the same 3 epochs, with a 2.0 learning rate, and a final training loss of 0.2085 (Figure 10).

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image7-c.png|centre|thumb|474x474px|'''Figure 10''': ChatGPT-4o Mini Fine Tuning (Top) & ChatGPT-4o Full Fine Tuning (Bottom). (Generated by Student, Adapted from OpenAI API)]]</div>

LlaMa-3.x; This model is instilled directly into the Hugging Face library and was downloaded into an external Standard Storage Device (SSD) due to the large model size and independence from conflicting environmental settings. The usage for research purposes was approved via request and requires specific API keys to utilize. Visual Studio Code directly hosted the environment for model memory and was loaded via transformers. AutoTokenizer.from_pretrained and transformers. AutoModelForCasualLM.from_pretrained. Inputs were loaded via CSV format and then made into a class compatible with the API format. This was combined with training specifications, with default learning rates, epoch counts, and the use of quantization to decrease model load on GPU and CPU usage.

Gemini 1.5 Pro; This model is fine-tuned through Google AI Studio API with a structured JSON similar to ChatGPT-4o. Key differences include more descriptive classifiers for parameters which are used to fine-tune the model. This file had to be uploaded to Google Cloud, tuned through vertexAI, and used gemini-1.5-pro-002 as the source model. This creates an endpoint through Google AI Studio which allows the model to be used. In all, roughly 7 models were fine-tuned per transcription, each with differing model configurations such as a test-train split or learning rate, with key formatting issues being the primary cause for model error and repeated creation. Almost all tuning specifications were provided via official guides, which held nondescript JSON formatting, tuning, and utilization example code. [40,41,42] The final model cost roughly ~1.50$, and was trained with the default epoch count of  5 and learning rates of 2.0. Each epoch made ~2.5k predictions and reached a final training loss of 1.98 (Figure 11).

[[Image:Draft_Petrizzo_769196690-image9-c.png|center|600px|thumb|'''Figure 11''': Gemini 1.5 Pro Fine Tuning (Generated by Student, Adapted from Google Cloud.)]]

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;"></div>

Transfer-Learning; Each model utilized identical transfer-learning specifications. Firstly, the transcription is loaded according to the selected patient. For each row in the transcription, the patient and therapist message is appended to a string which is then passed into the model as background/system information. The model used is the base-line version which was used for fine-tuning. A series of key background information is also passed, such as instruction to behave like the transcription and to respond as if this is a new therapist. For each sent message, the user-input and background information is passed in its entirety. Gemini 1.5 Pro struggled to differentiate transcription from user-input, in large part due to no way to differ system-content and user-content in chat completion. The Llama series cannot be transfer-learned as it has no prior trained information.

Customizable Model; A descriptable model which allows the user to input background information and a mental illness. This was created to mitigate the need for transcripts and for the user to practice with a patient of a select mental illness or select behaviors. However, these models cannot be compared with the cosine similarity due to a comparative transcript not being available. Instead, professional feedback is the only metric available to inform the quality of these models. Both GPT models are available as customizable models, Gemini and Llama are not for their aforementioned characteristics.

=== ''Usage'' ===

The GUI is split into two pages, one for initial model selection/customization, and the second to initiate the conversation (Figure 12). The initial page houses multiple patient selections alongside a ML model choice. Further the last selection allows the user to customize the model with a description and mental illness. The “Tuned” checkbox represents the decision between Tuned(Fine-Tuned) and unTuned(Transfer-Learning) models. Pressing submit redirects the user to the conversational page. This page has options to send messages through text or voice-recording, as well as buttons to save the conversation in both a CSV and video format for instructor evaluation.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Draft_Petrizzo_769196690-image4-c.png|centre|thumb|306x306px|'''Figure 12''': Usage. (Generated by Student)]]</div>

==RESULTS AND ANALYSIS==

The individual metrics for model accuracy, recall, and instruction following are characteristically different depending on the exact patient transcript utilized (Table 3). This could be due to varying data quality, which cannot be verified for patient anonymization. Likewise, data quality is not standardized nor verified for usage as a virtual patient. Each model received an accuracy through a randomized series of 50 messages from the training set, the messages were randomized with a random number generator. The randomized set utilized identical seeds to ensure model comparison.

=== ''Fine-Tuned Models'' ===

'''Table  3:''' Fine-Tuned Model Metrics. Generated by Student

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;vertical-align: top;"|Model
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o Mini''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o''

''Full''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''Gemini 1.5 Pro''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''LlaMa-3.2v-1B''
|-
|  rowspan='1' style="border-top: 1pt solid black;border-bottom: 1pt solid black;"|''Transcript 1''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|93.66
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|88.88
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|86.12
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|83.11
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Model Loss
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|0.12
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|0.21
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|-
|  style="text-align: center;vertical-align: top;"|1.73
|-
|  rowspan='1' style="border-top: 1pt solid black;border-bottom: 1pt solid black;"|''Transcript 2''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="text-align: center;vertical-align: top;"|92.31
|  style="text-align: center;vertical-align: top;"|86.80
|  style="text-align: center;vertical-align: top;"|85.18
|  style="text-align: center;vertical-align: top;"|83.68
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Model Loss
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|0.06
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|1.66
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|-
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|1.76
|-
|  rowspan='1' style="border-top: 1pt solid black;"|''Average''
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|92.99
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|87.84
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|85.65
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|83.40
|-
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Model Loss
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|0.09
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|0.93
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|-
|  style="border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|1.74
|}

Interestingly, ChatGPT-4o Mini performed higher statistically compared to the larger, superior ChatGPT-4o Full model. Both models were trained identically, utilizing the exact same script and training data besides changing the model specification. The training for both models seems to lack convergence of training loss, which is something that warrants further analysis and fine-tuning more than the recommended settings. This may show a superiority in generalized tasks through a smaller network, rather than an ambitiously large dataset. Further, the cost difference for training and usage makes ChatGPT-4o-mini much more applicable in business-end environments, as the testing cost was near negligible for ChatGPT-4o-mini while reaching upwards of 0.10$ for 50 inputs of ChatGPT-4o-full, showing a stark increase.

'''Table  4: '''Transfer-Learning Model Metrics. Generated by Student

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;vertical-align: top;"|Model
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o Mini''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''ChatGPT-4o''

''Full''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|''Gemini 1.5 Pro''
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;"|''Transcript 1''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|83.21
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|91.25
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|87.18
|-
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;"|''Transcript 2''
|  style="border-top: 1pt solid black;border-bottom: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="text-align: center;vertical-align: top;"|81.44
|  style="text-align: center;vertical-align: top;"|86.52
|  style="text-align: center;vertical-align: top;"|80.39
|-
|  style="border-top: 1pt solid black;"|''Average''
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|Cos. Sim.
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|82.33
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|88.88
|  style="border-top: 1pt solid black;text-align: center;vertical-align: top;"|83.79
|}

The accuracies in the transfer-learning models are similar to those of the Fine-tuned models. However, the GPT-4o-mini model performed worse in comparison to the larger model, and all of these models have seemingly lower accuracies than the Fine-tuned counterparts. Without professional feedback, it would be incredibly difficult to differentiate the quality of these models. However, some key differences are apparent, such as the transfer-learning models have a much more descriptive and longer output. These models are more deterministic, having closer to identical messages upon the same inputs but much more expensive costs. For comparison, the transfer-learning models were roughly 10x as expensive but had higher quality and more descriptive outputs which may explain semantically dissimilar messages.

The Llama-3.2v-1B was created as a comparative model to the larger and generally stronger GPT and Gemini series. This model achieved the lowest accuracy, which can be attributed to having the smallest model size when compared via parameter count to the other models. However, this model was released to be suitable for business environments with capabilities for fine-tuning accurately due to the model’s small size but retained capabilities to the 3.1 series. The training was completely customizable and allowed for much more selections compared to API-based models. The Llama-3.1v models were much more expensive than the 3.2v, and were virtually impossible to train on a singular device.

Gemini 1.5 Pro faced many challenges in tuning, documentation, testing, and model retention. The overall performance of Gemini 1.5 Pro is high, but the model lacked the ability to grasp the contents of small-sized inputs, even with accelerated learning rates. As such, the accuracy of the model is lower than ChatGPT, but a larger transcription may prove highly beneficial. The fine-tuned and transfer-learning model strongly represented the base model more than the patient transcription, having inbuilt descriptors to behave dissimilar to human-like characteristics. Lastly, the safety guidelines for Gemini 1.5 Pro were very aggressive, meaning most inputs ended up being denied inputs. This was with the safety guidelines specifically declared, warnings accepted, and all token counts as high as possible, but most inputs were still declined without a warning output. While not necessarily negative in general usage as a chatbot, the Gemini model failed as a virtual patient.

The ability for a model to respond quickly is also a vital component of a chatbot design, to enhance user satisfaction and make fluid conversation. Having the model entirely installed on the system makes Llama the fastest model, being dependent on the configuration and specification of the system, while Gemini 1.5 Pro took extensive safety measures; not necessarily negative, but would extend prediction times dramatically. Furthermore, the rate-limiting of API-based models, such as the ChatGPT and Gemini series used in this paper, may incidentally cause inputs to be buffered or lost entirely. The transfer-learning models responded much quicker than the equivalent fine-tuned models. Oftentimes being faster in comparison of milliseconds to multiple seconds.

==CONCLUSION==

''' '''Utilizing modern Artificial Intelligence models to develop virtual patients is a promising applicable concept which warrants further development with higher computational power. The need for such developments is high; especially in ensuring resident training quality and creating availability to simulated patients. Furthermore, the promise for such concepts to be successful increases with the development of more advanced models that can capture semantic relations with even fewer inputs. The creation of a website interface was successfully developed to be utilized; however, there are various hurdles before full deployment into professional environments.

=== ''Limitations'' ===

''' '''The limitations within this paper reflect those of computational power and time constraints. This is especially apparent in rate-limiting, training sizes, documentation difficulty, and data availability. The limitations in creating virtual patients specifically stem from data availability, where finding large amounts of data brings concerns to identification and the minimal data may reflect biases; especially as the transcripts were from the same source. However, potential transcriptions would best be hosted under a school of therapy or similar environment, where connections are awaiting a response. The training in a large model requires hundreds of hours of fine-tuning, which may then be nullified due to a simple error and then restarted, forcing more time to train. Amplified by novel documentation of model training and usage creates large gaps in public knowledge for utilizing these modern models. Furthermore, the limitations in artificial outputs make standardization of each model difficult; as the probabilistic nature of a Machine Learning model prevents consistent responses. Lastly, the testing from professionals is ongoing for these models, and may dramatically impact reflections about model accuracy and capacity for expansion. Connections to certified psychotherapists have been made, and feedback is awaited.

=== ''Future Research'' ===

As stated prior, the need for further evaluation and training adjustments is an ongoing and future focus of research. Notably, the Llama series requires more hours of fine-tuning to be evaluated for accuracy and then deployed. A continuation to evaluate models with psychotherapists and residents is also a focus for future research. More generally; additional time is necessary to develop the limitations and expand the potential to suit the unique challenge of replicating patients with specific characteristics. This could require a dataset of “unique” individuals each with their characteristic traits to warrant development. In addressing transcription needs, Generative Artificial Intelligence is a hesitant alternative that has high potential but demands broad exceptionally large models to be trained. The usage of generation to both create transcriptions and generate images representing the patient’s emotions has the potential to further reflect the nuances of in-person conversation, but requires further developments in generative AI.

=== ''Potential'' ===

There is potential apparent in each model’s displayed accuracy and further metrics to evaluate the model’s performance are necessary to determine their use in real-life application. For example; ChatGPT-4o mini having a higher accuracy than the base-model equivalent is highly counterintuitive and needs to be evaluated further. The accuracy of the models overall reflects some semantic correspondence successfully gathered by the model, which can also likely be improved through more data. Interaction with the model was successfully implemented and developed intentionally to provide the ability for model addition as technology develops further. The website was also developed with the ability to be converted into an app for mobile devices, further supporting scalability as models are evaluated further. By utilizing Electron as a secondary port, the website has executable files, a website interface, and mobile conversion capacities. This allows easy integration into ongoing training systems for residents. In all, this paper demonstrates the potential of AI to greatly improve the availability of resident training.

==ACKNOWLEDGEMENTS:==

In Loving Memory of Zuri Leung.

Thank you to Ms. Colette, Ms. Beatty, and Dr. Kramer for their amazing guidance.

==REFERENCES:==

<nowiki>[1] ACGME Competencies (2024). Accreditation Council for Graduate Medical Education. Retrieved on September 26, 2024, from URL: https://www.acgme.org/programs-and-institutions/programs/common-program-requirements/</nowiki>

[2] NYS Mental Health Counseling (2024). New York State Licensed Professionals. NYSED. Retrieved on September 26, 2024, from URL: https://www.op.nysed.gov/professions/mental-health-counselors/license-requirements#:~:text=Education%20Requirements,-To%20meet%20the&text=accredited%20as%20a%20clinical%20mental,another%20acceptable%20accrediting%20agency%3B%20or

[3] Barrows, H. S. (1993). An Overview of the Uses of Standardized Patients for Teaching and Evaluating Clinical Skills. Academic Medicine, 68(8), 443-451

[4] Lovink, A., Groenier, M., Niet, A., Miedama, H., and Rethans, J.J. (2024). How simulated patients contribute to student learning in an authentic way, an interview study. Advancing Simulation, 9(4), 1-10.

[5] Cleland, A. J., Abe, K. & Rethans. J, J. (2009). The use of simulated patients in medical education: AMEE Guide No 42. Medical Teacher, 31(6), 477-486

<nowiki>[6] Virtual Patients. Body Interact. Retrieved on September 27, 2024, from URL: https://bodyinteract.com/virtual-patient-simulator/</nowiki>

[7] Zalewski, B., Guziak, M., and Walkiewicz, M. (2023). Developing Simulated and Virtual Patients in Psychological Assessment - Method, Insights and Recommendations. PubMed, 12(1), 455-461.

[8] What is Machine Learning? (2024a). IBM. Retrieved on October 18, from URL: https://www.ibm.com/topics/machine-learning

[9] Turing, A. M. (1950). Computing Machinery and Intelligence. Oxford, 59(236), 433-460.

[10] ADAMW. (2023). PyTorch. Retrieved on October 22, 2024, from URL: https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html

[11] Stryker, C., and Holdsworth, J. (2024b). What is NLP? IBM. Retrieved on October 22, 2024, from URL: https://www.ibm.com/topics/natural-language-processing

[12] Lederer, J. (2021). Activation Functions in Artificial Neural Networks: A Systematic Overview. arXiv:2101.09957.

[13] Baheti, P. (2021). Activation Functions in Neural Networks [12 Types & Use Cases]. v7. Retrieved on October 22, 2024, from URL: https://www.v7labs.com/blog/neural-networks-activation-functions

[14] What is a Neural Network? (2024). Amazon Web Services. Retrieved on October 18, from URL: https://aws.amazon.com/what-is/neural-network/#:~:text=Artificial%20neural%20networks%20learn%20continuously,paths%20in%20the%20neural%20network.

[15] Stryker, C. (2024a). What is a recurrent neural network? IBM. Retrieved on October 18, 2024, from URL:  https://www.ibm.com/topics/recurrent-neural-networks

[16] Transformer’s Encoder-Decoder. 2021. KiKaBeN. Retrieved on October 18, 2024, from URL: https://kikaben.com/transformers-encoder-decoder/

[17] Ba, J. L., Kiros, J. R., and Hinton, G. E. (2016). Layer Normalization. arXiv:1607.0650.

[18,19] What is Mixture of Experts? (2024b). IBM. Retrieved on November 21, from URL: https://www.ibm.com/topics/mixture-of-experts

<nowiki>[20,21] Ahmed, A., Hayat, H., and Hayat, D. (2024). Comparing GPT-4o, LLaMA 3.1, and Claude 3.5 Sonnet. Walturn. Retrieved on September 23, 2024, from URL: https://www.walturn.com/insights/comparing-gpt-4o-llama-3-1-and-claude-3-5-sonnet</nowiki>

[22] Haywood, D. (2024). Gpt-4o tokens per second comparable to gpt-3.5-turbo. Data and analysis. OpenAI Developer Forum. Retrieved on September 23, 2023, from URL: https://community.openai.com/t/gpt-4o-tokens-per-second-comparable-to-gpt-3-5-turbo-data-and-analysis/768559

[23] Kilpatrick, L. and Mallick, B. S. (2024). Updated production-ready Gemini models, reduced 1.5 Pro pricing, increased rate limits, and more. Gemini. Retrieved on November 21, 2024, from https://developers.googleblog.com/en/updated-gemini-models-reduced-15-pro-pricing-increased-rate-limits-and-more/

[24] Pichai, S. and Hassabis, D. (2024). Our next-generation model: Gemini 1.5. Google Blog. Retrieved on November 21, 2024, from URL: https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/#sundar-note

[25] Zhou, J., Lu, T., Mishra, S., Brahma, S., Basu, S., Luan, Y., Zhou, D., and pHou, L. (2023). Instruction-Following Evaluation for Large Language Models. arXiv:2311.07911.

[26] Shaham, U., Ivgi, M., Efrat, A., Berant, J., and Levy, O. (2023). ZeroSCROLLS: A Zero-Shot Benchmark for Long Text Understanding. arXiv:2305.14196.

[27] Zhang, X., Chen, Y., Hu, S., Xu, Z., Chen, J., Hao, M. K., Han, X., … Sun, M. (2024). ∞Bench: Extending Long Context Evaluation Beyond 100K Tokens. arXiv:2402.13718.

[28] Wang, H., Shi, H., Tan, S., Qin, W., Wang, W., Zhang, T., Nambi, A., … Wang, H. (2024). Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models. arXiv:2406.11230.

<nowiki>[29] Gadesha, V. and Kavlakoglu, E. (2024). What is chain of thoughts (CoT)? IBM. Retrieved on October 24, 2024, from URL: https://www.ibm.com/topics/chain-of-thoughts#:~:text=Chain%20of%20thought%20prompting%20is,steps%20towards%20a%20final%20resolution.</nowiki>

[30] Ahmed, A., Hayat, H., and Hayat, D. (2024). Comparing GPT-4o, LLaMA 3.1, and Claude 3.5 Sonnet. Walturn. Retrieved on September 23, 2024, from URL: https://www.walturn.com/insights/comparing-gpt-4o-llama-3-1-and-claude-3-5-sonnet

<nowiki>[31] Gemini 1.5 Pro (Sep): API Provider Benchmarking % Analysis. (2024a). Artificial Analysis. Retrieved on November 21, 2024, from URL: https://artificialanalysis.ai/models/gemini-1-5-pro/providers</nowiki>

[32] Gemini 1.5 Pro (Sep): Quality, Performance % Price Analysis. (2024b). Artificial Analysis. Retrieved on November 21, 2024, from URL: https://artificialanalysis.ai/models/gemini-1-5-pro

<nowiki>[33] Prepare supervised fine-tuning data for Gemini models. (2024a). Google Cloud. Retrieved on November 22, from URL: https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini-supervised-tuning-prepare</nowiki>

[34] The AI detective: The Needle in a Haystack test and how Gemini 1.5 Pro solves it. (2024b). Google Cloud. Retrieved on November 28, 2024, from URL: https://cloud.google.com/blog/products/ai-machine-learning/the-needle-in-the-haystack-test-and-how-gemini-pro-solves-it

[35] Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. (2024). Google DeepMind. Retrieved on November 27, 2024, from URL: https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf

[36] Instruction-Following Evaluation. (2024). Scale AI. Retrieved on November 21, 2024, from URL: https://scale.com/leaderboard/instruction_following

[37] Ye, J., Chen, X., Xu, N., Zu, C., Shao, Z., Liu, S., Cui, Y. ... Huang, X. (2023). A Comprehensive Capability Analysis of GPT-3 and GPT-3.5 Series Models. arXiv preprint arXiv:2303.10420.

[38] Wang, R., Milani, S., Chiu, C. J., Zhi, J., Eack, M. S., Labrum, T., Murphy, M. S., Jones, N ... Chem. Z. Z. (2024). PATIENT-Ψ: Using Large Language Models to Simulate Patients for Training Mental Health Professionals.  arXiv preprint arXiv:2405.19660.

[39] Wu, M. and Fishman, S. (2023). Data preparation and analysis for chat model fine-tuning.  OpenAI. Retrieved on October 17, 2024, from URL: https://cookbook.openai.com/examples/chat_finetuning_data_prep

[40] Fine-tuning tutorial. (2024). Google AI. Retrieved on November 23, 2024, from URL: https://ai.google.dev/gemini-api/docs/model-tuning/tutorial?lang=python

<nowiki>[41] Prepare supervised fine-tuning data for Gemini models. (2024a). Google Cloud. Retrieved on November 22, from URL: https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini-supervised-tuning-prepare</nowiki>

<nowiki>[42] Tune Gemini models by using supervised fine-tuning. (2024c). Google Cloud. Retrieved on November 22, from URL: https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini-use-supervised-tuning</nowiki>